Home
Deep neural network model introducing new novel matching layer called 'Normalized correlation' layer. This repository contain information about the datasets used, implementation code. The paper titled "Deep Neural Networks with Inexact Matching for Person Re-Identification" is accepted in NIPS-2016. You can find the paper here
Person re-identification is the task of matching a person's image across multiple camera views. It poses many challenges such as illumination variation, pose/viewpoint change, partial occlusion, low image quality etc.,
In this work, we propose a novel matching using "Normalized correlation" and "Wider search space" in Deep learning framework. The proposed method is more robust to Illumination variation (with the help of Normalized correlation), pose/viewpoint variation (owing to wider search carried out around the neighborhood of a pixel) and partial occlusion (with the combination of Normalized correlation and Wider search).
Our method gives better performance than the current state-of-the-art method in the large dataset CUHK03 (labeled and detected), mid-sized dataset CUHK01 (single-shot test100 and test486 settings) and small dataset QMULGRID. Nevertheless, In our analysis, our method gives some occasional false matches (we owe it to the fact that we search in a wider space and normalized correlation shall match with multiple similar textures, as well as background). To overcome these false matches, we fuse our novel matching layer with Ahmed et al.(CVPR 2015)'s CrossInput Neighborhood layer, as it performs an exact matching and we hypothesize that it will consider the exact color/pixel value while matching and will make the model to learn complementary features along with inexact matching.